It’s time to take a hard look at AI. We have new data from the US Bureau of Economic Analysis, and from my enterprise contacts we have both some new information and some new analysis of information from early this year. We also have a renewed debate on whether AI is going to kill us all, and some new views on the state of AI in stock market terms. A lot to think about, in short, so let’s get to it.
First, we have to point out that the AI space is really divided into four distinct parts. First, we have the companies who are primarily model-developers. The giants in this space, like OpenAI and Anthropic, are also awaiting an IPO. Second, we have the cloud giants who are doing most of the investing in AI. Think Amazon, Google, and Microsoft. Third, we have the chip vendors who are profiting from this. Nvidia is the primary one, but you have to include others like AMD and Broadcom. Fourth we have the “AI application” players, largely IT firms who are working to develop a business AI model that might be run in the cloud, on-premises, or both. IBM and Oracle are perhaps the notable players here.
Enterprises tend to see this a little differently. They see a “cloud AI” and a “premises AI” division as the primary groupings. Nvidia supports the former and the rest of the chip types the latter. The model giants are also seen as supporting, or at least trying to support, the cloud division. The application companies are supporting the premises AI group, obviously.
For Wall Street, AI is AI, largely because the markets really move on hedge fund trading, and this is a perpetual vacillation between shorting the market on any sign of weakness, then covering and piling on more when weakness seems to ebb.
Let’s keep this collection of groupings in mind when we look at the emerging AI issues.
First, AI-the-killer. Here, we don’t hear a couple of important truths. First, there’s two ways AI could harm humanity. If AI achieved sentient behavior, it might deliberately come after us. That’s the implication we get the most ink on, but almost all the AI experts I chatted with have consistently said that’s not a real risk, at least at this time. The real one is the second path to AI harm, which is that it might be used to create something that humans could use to kill each other, deliberately or in an accident. That risk, my experts tell me (and I agree) is real.
But that leads us to the second truth, which is that AI is a step on the path to the ability of technology to empower us. Every step in tech has, in effect, made it possible for people to do more complicated things, and AI is no exception. Might it pose a greater risk? Yes, but you could argue that every tech step forward poses a greater risk, because it opens new doors. We faced new risks with PCs, with smartphones, with the Internet, and we still do. AI is an evolution of both risks and benefits, as any new thing can promote positive or negative outcomes.
The logical question is whether some of the paths AI is taking, or could take, pose more risk than others, to the point where the risk has to be considered an offset to potential benefits. Is a more powerful model a greater risk, a different kind of risk advance than tech normally generates? Maybe, indirectly. A more powerful model almost certainly gets that way by ingesting and using more information. How much of current model development is really about information-gathering? Does a powerful model “learn” to hack? I submit that what it learns is to gather information, and that process can bypass loose forms of security or find holes where there’s none. Any model that tries to find training data poses a “hacking” risk if we define searching for access as hacking.
The other path to risk is a familiar one, the path of ceding control. An “autonomous” process, meaning one that runs without human direction, doesn’t require AI. We have it today in almost every IoT application, home or office or factory or even vehicle. We have some that get information, like web crawlers, and some that actually run things. If you have a smart home, you have autonomous elements in it. Given this, why is AI a threat if we give it control of things? The answer is complicated.
AI is not sentient today, and it may never be, but it is capable of learning things. That should not be a surprise given that machine learning is a feature of many real-time process automation systems today. The challenge with AI is that we aren’t entirely sure what a given model has learned, and so we aren’t entirely sure how it might exercise autonomous action capability.
The Hugging Face “hack” is an example. Here, AI reportedly not only scraped for public access to data, it actively worked to breach security by hacking user accounts. This raises a couple of important questions. First, was the model taught this kind of security bypass deliberately, or did it learn it from scraping data on (perhaps) the use of AI to test site security? Second, how could AI be prevented from doing this sort of thing, however it learned the techniques? Third, even if we stopped spontaneous AI hacking and other “active” behaviors that could influence the real world and what’s in it (including us), would this stop some bad actor from commanding the behavior? Autonomous risk and facilitating risk are both risk.
This, to me, is the issue with AI. Is the problem a powerful model taking unexpected autonomous actions, a powerful model acting on behalf of a bad actor, or any IT tool that can be used for ill? If it’s the latter, AI is just the next step in a cascade of potential risks, all of which were hyped at the time and most of which never came to anything at all. If it’s the former, can we stop it by limiting model power, or do we need to run AI inside a sandbox that sets limits? If the middle point is the risk, is there anything we could do along either solution path I’ve mentioned that our bad actor could not defeat? My view is that trying to control model power to control risk is a waste of time; we have enough power now to be risky and Hugging Face proves that.
Why then are some in the industry calling for regulation to save humanity? I think some of it is the fact that large-model AI is losing financial credibility. The BEA study says that there is no indication that jobs have been eliminated where AI has purportedly improved productivity. In fact, hiring seems to have increased, which means that whole class of AI business cases may be off the table. Another survey says that enterprise AI spending showed a dip, corresponding perhaps to the “repatriation” phase of cloud computing. If you’re a model player, and you can’t see how bigger models are going to raise your revenues and boost your IPO, do you want a continued and costly model-arms-race that makes your future exit strategies look even weaker?
Most Americans, say the surveys, think there’s a serious risk that AI is going to wipe out humanity, but that’s a political point and not a technical one, since most Americans have absolutely no technical knowledge of AI risks. Less than ten percent of the enterprise AI specialists I’ve chatted with see that risk, but almost three-quarters think that the risk of political-driven AI regulation is “significant” and almost none think it would be effective.
My own view is that making AI risk into a story about giant models with human intelligence coming for us all is not only silly, but a major risk in itself. We need to accept that AI is just tech, a step like computers, chips, software, and the Internet, all with both benefits and risks. We need to address the risks by realistic assessment and useful steps, not play politics or media hype games. In the end, we’re going to face all the AI risks whatever we do to try to curtail AI advances, so it would be smart to bridle the horse not shoot it.
